Forecasting portfolio variance: a new decomposition approach

This paper proposes a new decomposition approach by separating realized covariation into components based on signs (positive and negative) and magnitudes (continuous, small jump, and large jump). The motivation behind this decomposition is that certain variation components can be more useful in forecasting than others. Including only “information-rich” components in realized (co)-variation forecasting models can improve predictive accuracy. Using various machine learning models, the marginal predictive content of each variation components can be assessed. The empirical exercise is based on all constituent of S &P 500 stocks between 2010 and 2019. We find that standard machine learning methods without the more granular variation measures offer limited improvement to out-of-sample fit (likely due to low signal-to-noise ratios) relative to benchmark HAR-type forecasting models. However, sparse models, which are specified using predictors selected using a first “variable selection” yield significant improvements in predictive accuracy when the decomposed variation measures are included. These predictive gains can be traced to the identification of short-lived pricing signals associated with co-jumps.